podcast-ad
Builds converting podcast-style ads in Palmier Pro as an overheard two-person conversation — sceptic arc, believer mechanism, opposed reverse-shot stills, locked voice + eyeline, exact off-screen backchannels, Flux-first beats, assemble + silence trim. Use for podcast ads, UGC conversation ads, sceptic-to-believer product spots — not for single-hook talking-head ads (use ugc-video-prompts).
palmier-io · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add palmier-io/palmier-skills --skill podcast-ad技能说明
name: podcast-ad description: Builds converting podcast-style ads in Palmier Pro as an overheard two-person conversation — sceptic arc, believer mechanism, opposed reverse-shot stills, locked voice + eyeline, exact off-screen backchannels, Flux-first beats, assemble + silence trim. Use for podcast ads, UGC conversation ads, sceptic-to-believer product spots — not for single-hook talking-head ads (use ugc-video-prompts).
Podcast Ad
The ad is two people, one room, one conversation. Viewer overhears; nobody pitches the lens. Script is ~80% of the result.
Palmier does not generate a full multi-speaker room in one pass — segment the exchange into model-length beats, lock room/identity with matching reverse-shot stills, then assemble on the timeline.
Pair with video-prompting. For faces that must match brand talent across variants, run character-pipeline first. Finish captions/layout with ugc-editing patterns when helpful.
Why this format works
- She never addresses camera — kills “ad argument in the head”
- His doubts are the buyer’s doubts — conversion when he clicks
- Pattern-matches podcast clips in the feed, not classic ad shapes
Break overheard staging (lens eye contact, hard CTA, cheerleading sceptic) and you lose the format.
Step 1 — Write the exchange (sceptic first)
| Role | Job |
|---|---|
| Believer | Transformation. Casual certainty. Talks to a mate, not a customer. |
| Sceptic | Doubt → convert on camera. Write him first — his arc is the ad. |
Sceptic’s four beats (required)
- Open doubt — names the quiet part out loud
- Honest question — hands her the floor without agreeing
- The click — mechanism connects to his life (not a parrot of her claim)
- Concession — stops arguing; does not gush
Believer craft: human comparison over feature list; disarm the scary version, then educate with a specific mechanism; callback close; no hard CTA inside the dialogue.
Map the script onto generation beats that fit list_models durations (often 4–10s). Prefer tight durations (often 5–6s). One conversational turn (or half-turn) per clip is safer than cramming. Flux 3 minimum is 5s.
Backchannels (reactive ad-libs)
Agent judgment (this skill): on longer believer turns, decide whether 1–2 quiet off-screen listener murmurs earn a place — sparse, woven through her speech (not parked only at the end), not after every clause, not a reverse-shot cutaway. Skip short callbacks and all sceptic-speaking beats.
Model prompt: never paste that judgment as a menu (uh-huh / oh? / mm). After you decide, write exact off-screen lines and where they sit relative to her words.
| Agent decides | Prompt must say |
|---|---|
| Whether this beat gets a backchannel | Exact phrase(s): "Oh?" |
| Which 1–2 murmurs + placement | After which of her clauses |
| Same sceptic Voice lock, quieter than her | Verbatim Voice + “quieter than her, audio only, not on camera” |
Example — agent chose two spots; prompt gets:
[DIALOGUE]
Believer says exactly: "Those water-retention drops changed everything —
one salty meal prep day and I'd look puffy. Now I just don't."
Voice: <believer Voice lock verbatim>
[BACKCHANNEL]
Off-screen male partner only (not on camera). Same sceptic Voice lock, quieter than her.
After she says "changed everything" he says exactly: "Oh?"
After she says "I'd look puffy" he says exactly: "Really?"
No other reactions. Do not invent a second person in frame.
Step 2 — Two matching start frames
list_models({ type: "image" }) → generate_image for each speaker.
Still models (validated)
| Prefer | Model id | Notes |
|---|---|---|
| 1st | grok-imagine (Grok Imagine 2) | Strong photoreal + natural podcast staging; no referenceMediaRefs in Palmier yet |
| 1st (tie) | gpt-image-2 (GPT Image 2) | Use for reverse still with refs; resolution: "1024x1536" for 9:16 |
| 1st (tie) | seedream-v5-pro (Seedream 5.0 Pro) | Refs OK; aspect often portrait_16_9 |
Do not default to Nano Banana / Krea / Ideogram / Recraft / MAI / Reve for podcast stills unless asked or the preferred trio fails.
These stills are shot / reverse-shot anchors: same room + camera grammar, opposed eyelines.
Hard rules
- One person only. No foreground shoulder, OTS, interviewer blur, or second body.
- Same camera grammar on both. Height, distance (chest-up), table edge, wall, lights — copy-paste. Only wardrobe + face + gesture + eyeline side change.
- Same room word-for-word.
- Flip the reverse shot. Believer frame-RIGHT ↔ sceptic frame-LEFT (or vice versa). Mirror mic/boom side with the subject.
- Generate A, then B with
referenceMediaRefs: [stillA](use GPT Image 2 / Seedream if A was Grok): preserve room + camera height/distance — replace subject and flip eyeline.
Prompt pattern
Podcast studio mid-shot, chest-up, eye-level camera, locked framing.
Single subject only — no other people, no foreground shoulder, no over-the-shoulder.
Wooden table edge low in frame, broadcast mic on boom in shot.
Background: <identical room description>.
Warm practicals, normal exposure, real room — not a film set.
Subject mid-gesture, mouth slightly open — not into lens.
Then only subject + eyeline differ:
Believer: <wardrobe/face>. Eyeline: looking frame-RIGHT toward off-screen partner.
Sceptic: <wardrobe/face>. Eyeline: looking frame-LEFT (OPPOSITE of believer).
Preserve room and camera height/distance from reference — flip orientation only.
Negatives: no second person, no OTS, no interviewer, no looking into lens, no same-side eyeline as the other speaker.
Frame checklist
- One person only; no OTS / shoulder
- Matching height, distance, room
- Mic in front of speaker
- Off-axis eyeline; opposed across the pair
- Mid-gesture / mouth slightly o